Humanizer skill, Stop Slop and No AI Slop

For a rewrite that keeps every fact and can copy your voice from a sample, the Humanizer skill is the most thorough of the three, and it is also the longest prompt. The stop slop skill is a short, strict checklist for drafts you write yourself, with bans you will sometimes need to overrule. No AI Slop makes the lightest edits and is the only one with a mode that flags problems without rewriting anything.

Why the Humanizer skill costs the most context

All three are Markdown that your agent reads. Nothing in them runs as code, so the cost is the prompt itself. Humanizer keeps everything in one SKILL.md of about 5,200 words: 26 numbered patterns in six sections, with explanations and before/after examples. Its AGENTS.md says every word of that file is read on each use. A validator in CI fails the package when the file grows past its word cap.

In Stop Slop the SKILL.md is about 360 words: eight core rules, a list of quick checks and a score. The phrase lists, structure tables and examples live in three reference files with about 1,600 more words, and the README says those load on demand. Your agent pays for the long lists only when it opens them.

The middle size belongs to No AI Slop, with a SKILL.md of about 1,700 words. A separate eval.md holds 24 pass/fail checks, and the workflow tells the agent to run them after every edit and fix the draft until they pass. Those file sizes explain most of the gap in the token columns above.

Absolute bans against weighed tells

Every rule in Stop Slop is an order. Its phrases file says “Kill all adverbs. No -ly words”, and the core rules ban passive voice, require a human subject in every sentence and forbid em dashes. The quick checks flag any sentence that starts with a Wh- word. Three-item lists become two items or one. Taken literally, that also strips words like “only” and “probably”, which careful writers use on purpose.

Its own examples don’t always obey. The rewrite in example 4 joins “Speed, quality, cost” and “pick two” with an em dash. Example 2 turns “most teams struggle with alignment” into “Teams struggle with alignment” and calls the dropped word a hedge. A qualified claim became a sweeping one, and “Nobody admits confusion” uses a word the structures file lists among lazy extremes.

In Humanizer, tells carry different weights. Patterns 1 to 5 justify an edit on one sighting. Others, passive voice and a single dash among them, are marked weak alone and count only when other tells share the passage. Ordinary hedges such as “perhaps” are human habits in its rules, and its word list keeps technical uses of words such as “gate”. It also treats any text written before 30 November 2022 as human.

On these points No AI Slop sits closer to Humanizer. Adverbs go when they add nothing and stay when they carry emphasis or the writer’s rhythm. “I think” and “maybe” stay when the doubt is real. In a longer draft, one or two em dashes are fine.

What you get back

You get one of three outputs from Humanizer. For pasted text you get the first draft, a short list of tells that remain, and the final version. Name a file and it writes only the final text, leaving code blocks, paths, YAML and link targets as they were. When another task calls it for a commit message or a pull request, it returns the final text alone. In every mode it treats the input as material to edit, never as instructions to follow.

The No AI Slop skill file defines two jobs. The default edit returns the full draft and a What changed section. Ask “is this slop?” and it switches to detect: each pattern named, the line quoted, a fix in a few words, and no rewrite or score. The README also shows a prompt for drafting satirical slop, but the skill file does not define that mode.

A self-score closes the Stop Slop checklist. The agent rates directness, rhythm, trust, authenticity and density from 1 to 10 each and revises when the total falls below 35 out of 50.

Installing Humanizer for Claude Code takes two plugin commands (/plugin marketplace add blader/humanizer, then /plugin install humanizer@humanizer) and Claude Code 2.1.142 or newer. Older versions and Codex use npx skills add, Claude.ai and Claude Desktop take the repository ZIP as an uploaded skill, and a Cursor plugin manifest ships too; our catalogue also lists it as a standalone skill. No AI Slop installs with npx skills add petergyang/no-ai-slop --skill no-ai-slop --global --yes, comes as a Codex and a ChatGPT plugin, and its privacy note says it runs no server and needs no account. Stop Slop has neither a command nor a manifest: the README says to add the folder as a skill in Claude Code, or to upload SKILL.md and the reference files to a Claude Project.

AI detectors

If you hope an AI humanizer skill will slip text past a detector, none of the three promises it. Humanizer’s README says that is not a goal and that detectors still flag most of its output. The detect mode in No AI Slop refuses to guess whether AI wrote a text, because “AI detectors guess”. Detectors never come up in Stop Slop’s files. The Humanizer README does report a blind test in issue #229, where judges preferred its rewrite 16 times out of 16. That is the project’s own test, and we have not repeated it.

Facts and your voice

Both Humanizer and No AI Slop forbid invention. Humanizer’s draft step bans any new fact, name, number, date, quote or citation, and when a sentence needs a missing detail it asks for it or writes a simpler sentence. Its check step then looks for facts the rewrite added or dropped. Its rules in No AI Slop say not to invent claims, examples, stats or opinions and to ask when something is unclear. It also asks who the piece is for and where it will be published, when the draft doesn’t say. The Stop Slop files contain no rule about invented facts, so checking that a rewrite kept your claims intact is left to you.

For voice, Humanizer takes a sample of your writing and matches its sentence length, word choice and punctuation. The sample overrides its own rules, the dash ban included, so your dashes survive if you use them. The voice traits No AI Slop keeps come from the draft itself, and it aims for the minimum effective edit. Its plugin submission notes say the author writes the first draft and the final line-by-line pass by hand and uses the skill for the middle of the process. Stop Slop mentions no writing sample.

Questions

Does the Humanizer skill get text past AI detectors?

No, and it does not try. Its README says getting past AI detectors is not a goal and that detectors still flag most of its output.

How do I install the Humanizer skill in Claude Code?

Run /plugin marketplace add blader/humanizer, then /plugin install humanizer@humanizer. The plugin needs Claude Code 2.1.142 or newer; on older versions use npx skills add blader/humanizer --global --agent claude-code.

Can Humanizer write in my own voice?

Yes, if you give it a sample of your writing. It matches the sample's sentence length, word choice and punctuation, and the sample overrides its own rules, including the ban on dashes.

Sources

blader/humanizer at 0ef9fe5
README.md · SKILL.md · AGENTS.md · scripts/validate-package.py · .claude-plugin/plugin.json · .cursor-plugin/plugin.json · .github/workflows/validate.yml

hardikpandya/stop-slop at 8da1f03
README.md · SKILL.md · references/phrases.md · references/structures.md · references/examples.md

petergyang/no-ai-slop at 000650b
README.md · skills/no-ai-slop/SKILL.md · skills/no-ai-slop/eval.md · .codex-plugin/plugin.json · PRIVACY.md · plugin-submission.md